Self-attention-based Human Activity Detection Using Wearable Sensors

Abstract

Analyzing human activity through wearable sensors can assist applications connected to context—vigilance and health care. The proposed approach utilizes convolutional and recurrent modeling to express the space–time-dependent constraints from information provided by numerous sensors. We construct a neural network layout, a combination of convolutional neural network and recurrent neural network, which focuses on spatiotemporal characteristics provided by sensor time series data with short listing and learning crucial points by executing a self-attentive procedure. We exhibit validation of our designed strategy on the WISDM dataset and indicate that the self-attention technique achieves a remarkable amendment in presentation over deep neural networks through the union consisting of convolutional and recurrent network architectures. We also present that our designed algorithm offers a statistic performance improvement over other designed approaches for the information (dataset) under consideration. The presented technique allows precise interpretation of activity from numerous body part sensors through various time intervals. We evaluated our technique through various classification methods namely: MLP Classifier, gradient boosting, Random Forest Classifier, Conv2D–LSTM, and self-attention mechanism (proposed). The simulation results show that our proposed method achieved the highest accuracy of 89.58% out of all other methods. We compared the effects of the designed architecture scheme with a baseline constructed through ConvNN and RecNN models, and with earlier designed approaches. We figured out that our self-attention model outperformed the baseline model with high difference among all the considered metrics.